Minimize the Memory Dependencies for Loop Pipelining
Intel® oneAPI DPC++/C++ Compiler ensures that the memory accesses from the same thread respects the program order. When you compile an NDRange kernel, use barriers to synchronize memory accesses across threads in the same workgroup.
Loop dependencies might introduce bottlenecks for single work-item kernels due to latency associated with the memory accesses. The Intel® oneAPI DPC++/C++ Compiler defers a memory operation until a dependent memory operation completes. This could affect the loop initiation interval (II). The Intel® oneAPI DPC++/C++ Compiler indicates the memory dependencies in the optimization report.
To minimize the impact of memory dependencies for loop pipelining:
- Ensure that the Intel® oneAPI DPC++/C++ Compiler does not assume false dependencies.
- When the static memory dependence analysis fails to prove that dependency does not exist, the Intel® oneAPI DPC++/C++ Compiler assumes that a dependency exists and modifies the kernel execution to enforce the dependency. The impact of the dependency enforcement is lower if the memory system is stall-free.
- Write-after-read operations with data dependency on a load-store unit can take just two clock cycles (II=2). Other stall-free scenarios can take up to seven clock cycles.
- The Intel® oneAPI DPC++/C++ Compiler can fully resolve the read-after-write (control dependency) operation.
- Override the static memory dependence analysis by adding the line [[intel::ivdep]] before the loop in your kernel code if you are sure that it carries no dependencies. For more information, refer to ivdep Attribute